Sound is a powerful source of information, capable of revealing what people are doing, from cooking to vacuuming. However, the use of audio sensors in smart devices poses significant privacy risks, as microphones can capture sensitive conversations. To address this issue, researchers at Carnegie Mellon University (CMU) have developed Kirigami, an on-device filter designed to detect and delete human speech segments before the data is used for activity recognition, ensuring privacy while maintaining the utility of the audio. Sudershan Boovaraghavan, a Ph.D. graduate from CMU's Software and Societal Systems Department (S3D), emphasizes the dual nature of sound data. "While sound can power valuable applications like health monitoring and environmental sensing," he said, "it can also be used to invade people's privacy." Kirigami offers a solution by being installed on various sensors with microphones, filtering out speech before the data is transmitted. Traditional privacy-preserving methods, such as excluding certain frequencies or training computers to ignore speech, have limitations. Generative AI models like OpenAI’s Whisper can reconstruct snippets of conversations from processed audio, making these methods less effective. Yuvraj Agarwal, an associate professor in S3D, the Human-Computer Interaction Institute (HCII), and the Electrical and Computer Engineering Department, explains, "The vast amount of data these models have can help them piece together fragments that were previously indecipherable. Kirigami prevents these models from accessing such snippets altogether." The team aimed to create a lightweight filter that could run on small, affordable microcontrollers, ensuring it could be used in a wide range of devices. The filter functions as a binary classifier, identifying and removing speech content. They designed it by analyzing the speech content recognition rate from deep-learning models. Kirigami can be configured to balance the aggressiveness of speech removal with the preservation of useful ambient sounds. Haozhe Zhou, an S3D doctoral student, noted, "Kirigami cuts out most of the speech content but not the other ambient sounds that are important for activity recognition. It can be used in conjunction with other techniques to enhance privacy further." The researchers are exploring numerous applications for activity sensing, including dementia care, monitoring children with ADHD, and assessing students for signs of depression. Mayank Goel, an associate professor in S3D and HCII, highlighted the non-invasive nature of these applications. "These scenarios require data from the person's daily life without infringing on their privacy," he said. "Kirigami helps achieve that balance." As the adoption of smart home infrastructure and the Internet of Things (IoT) continues to expand, the team believes developers can easily customize Kirigami to meet their specific privacy needs. The research, which has been published in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies and ACM MobiCom '24, demonstrates the potential of Kirigami to advance privacy-preserving technologies in a way that does not compromise the utility of smart devices. Industry Evaluation and Company Profiles: Industry insiders praise Kirigami as a significant advancement in balancing privacy and utility in smart devices. The lightweight design and adaptability make it a practical solution for a wide range of IoT applications. Carnegie Mellon University, known for its pioneering research in computer science and engineering, has once again demonstrated its commitment to addressing real-world challenges with innovative technology. The collaboration between multiple departments at CMU, including S3D, HCII, and the Electrical and Computer Engineering Department, reflects the interdisciplinary approach necessary to tackle complex issues in the tech landscape.